97% of organizations are building agentic AI. 12% run it through anything resembling a control plane. The gap isn't appetite for governance — it's that most enterprises have nowhere to put one.
Almost every enterprise conversation about governing AI agents is conducted in the vocabulary of policy. Steering committees. Acceptable-use standards. Model registries. Risk tiers borrowed from the EU AI Act and stapled to an internal wiki. All of it reasonable, and almost none of it load-bearing, because a policy is only as real as the place it gets enforced.
That place is missing in most enterprises, and the numbers now say so plainly. In a survey of 1,879 IT leaders published in April 2026, 97 percent of organizations reported exploring agentic AI strategies and 49 percent described their capability as advanced or expert. Thirty-six percent said they had a centralized approach to governance. Twelve percent actually used a centralized platform to exercise control.
The interesting number is not the 12. It is the 24-point drop between having a centralized approach and having a centralized mechanism — the distance between a governance model that exists on a slide and one that exists in a request path.
The reason isn't negligence. It's architecture.
It would be easy, and wrong, to read this as enterprises not caring. Ninety-four percent acknowledged that AI sprawl increases complexity and security risk. The constraint they describe is structural: 40 percent named legacy system fragmentation as their top barrier, 38 percent identified legacy systems as the primary reason agent projects stalled outright, and 48 percent said legacy integration was the single most critical capability they needed next.
Ninety-six percent said a unified platform for applications and agents was important. Seven percent had adopted one.
Read those together and the picture is not an organization that declined to govern. It is an organization that cannot, because governance is an enforcement function and enforcement requires a chokepoint. If an agent can reach a system of record through six different integration paths, four of which predate the AI program and two of which are undocumented, then the policy governing that agent is documentation, not control.
Governance without a routing layer is a filing system. You can write the rule, publish the rule, train people on the rule — and still have no place to apply it at the moment an agent takes an action.
Four layers, only one of which enforces anything
The confusion is worth pulling apart, because organizations routinely report themselves as "governed" on the strength of artifacts produced at the top two layers.
Visual 1 — Where agent governance is claimed vs. where it actually binds
Layer | What exists here | What it can actually stop |
|---|---|---|
Policy | Acceptable-use standards, risk tiers, approval committees | Nothing, directly. It defines intent and creates auditable expectations. |
Inventory | Agent registry, model cards, ownership records | Nothing, directly — but you cannot revoke what you cannot enumerate. |
Identity & entitlement | Machine identities, scoped credentials, per-agent permissions | Access an agent was never granted. Not misuse of access it was. |
Runtime routing | Gateway, broker or orchestration layer every agent call traverses | The action itself — rate, scope, data egress, human checkpoint, kill switch. |
How to read it: The 36 percent claiming a centralized governance approach are largely describing the top two rows. The 12 percent with a centralized control platform are the ones who have built the fourth. Most enterprise risk lives in the gap between them.
This also explains a finding that otherwise looks like squeamishness. Sixty-six percent of IT leaders said building human-in-the-loop checkpoints was technically difficult. That is not a statement about risk appetite. A checkpoint has to interrupt a call, hold state, route to a human, and resume — which is straightforward if every agent action passes through one broker, and close to unbuildable if agents call twelve systems directly through bespoke integrations. The organizations finding it hard are telling you exactly where their architecture is.
Trust is rising faster than the ability to enforce it
Meanwhile the appetite keeps climbing. Seventy-three percent of IT leaders now report high or moderate trust in autonomous agents, up sharply from the 40 percent who expressed comparable trust in generative AI writing code a year earlier. Financial services — the sector with the most to lose — leads on complete autonomous trust at 12 percent.
Trust curves are supposed to trail control curves. When they cross, the organization is not making a risk decision; it is making a risk assumption. And the 41 percent who rely on project-level rules rather than an enterprise framework have guaranteed that the assumption will be made independently, and differently, by every team that ships an agent.
That is the part worth naming out loud for a CIO. Project-level governance is not a lighter version of enterprise governance. It is a commitment to permanent drift: eleven teams, eleven interpretations, no single place to change the rule when the regulator, the auditor or the incident requires you to.
Three questions that separate governed from documented
Where does an agent's action get authorized? Name the component. If the answer is "in the agent's code" or "it depends on the team," you have policy, not governance.
Can you enumerate every agent in production today, with an owner? Not every approved agent. Every running one. The two lists diverge fast in organizations relying on project-level rules.
Can you revoke one agent's access to one system, everywhere, in one action? If revocation requires touching each integration, your blast radius during an incident is defined by your integration sprawl, not your policy.
What this changes
The practical implication is uncomfortable for anyone who scoped agent governance as a policy workstream with a compliance owner. The binding constraint sits with platform engineering and integration architecture, and it is the same constraint that has been stalling the projects themselves — 38 percent of them, by the survey's count. That is a rare alignment, and it is the argument to make internally: the integration layer that unblocks agent delivery is the same layer that makes agent governance enforceable. It is one investment, not two, and it is currently being pitched as neither.
The organizations that will look prudent in eighteen months are not the ones with the best-written AI policy. They are the ones that spent 2026 building a place to put it.
Sources and method. A BusinessInfomatics original. All survey figures — 97 percent exploring agentic AI strategies; 49 percent self-describing as advanced or expert; 36 percent with a centralized governance approach against 12 percent using a centralized control platform; 94 percent acknowledging AI sprawl risk; 66 percent finding human-in-the-loop checkpoints technically difficult; 41 percent relying on project-level rules; 40 percent citing legacy fragmentation as top barrier; 38 percent citing legacy systems as the reason projects stalled; 48 percent naming legacy integration as the most critical capability; 96 percent valuing a unified platform against 7 percent adopting one; and the trust figures of 73 percent and 12 percent — are from the OutSystems 2026 State of AI Development report, a survey of 1,879 IT leaders published April 13, 2026, as reported by TechHQ. The four-layer enforcement model and the three diagnostic questions are BusinessInfomatics' own analysis. Figures are dated as stated and not independently re-verified beyond the cited report.
